How does a convolutional code differ from a block code such as Reed-Solomon?

A block code partitions the data into fixed-length blocks and adds parity to each block independently, so encoding has no memory between blocks. A convolutional code has memory: each output depends on the current input and the contents of a shift register holding recent bits, so the encoder operates continuously on a stream rather than block by block. Convolutional codes excel against random bit errors on an additive-white-Gaussian-noise channel and pair naturally with soft-decision Viterbi decoding, which recovers 2 dB more gain than hard-decision. Reed-Solomon block codes excel against burst errors. Concatenated systems combine both: an outer Reed-Solomon (255,223) code corrects the residual bursts left by an inner convolutional code, the classic deep-space and DVB arrangement.

What coding gain does a rate-1/2 K=7 convolutional code provide at a BER of 1e-5?

With soft-decision Viterbi decoding on an AWGN channel, the standard rate-1/2, K=7 convolutional code delivers roughly 5.2 dB of coding gain at a bit error rate of 1e-5, reducing the required Eb/N0 from about 9.6 dB (uncoded BPSK) to about 4.4 dB. Hard-decision decoding gives about 2 dB less gain. Puncturing the code to rate 3/4 trades roughly 1.5 dB of that gain for higher throughput. The asymptotic coding gain is approximately 10*log10(R*dfree), where dfree=10 for this code, but real systems fall short of the asymptote at moderate BER because of the bit-error multiplicity associated with the free-distance error events.

Signal Processing

Convolutional Code

Q: What do constraint length and code rate mean for a convolutional code?

The constraint length K is the number of bit positions, including the current input, that influence each set of encoder outputs; it equals the number of shift-register stages plus one, so K=7 implies a 6-stage register and 64 trellis states (2^(K-1)). The code rate R = k/n is the ratio of input bits to output bits per step. A rate-1/2 encoder produces two coded bits for every information bit, doubling the transmitted bandwidth. Larger K increases the minimum free distance and therefore coding gain, but Viterbi decoder complexity grows as 2^(K-1), so most practical RF systems use K between 5 and 9. The canonical CCSDS and IEEE 802.11a code is rate-1/2, K=7, with generator polynomials 171 and 133 in octal.

/kon-vuh-LOO-shuh-nuhl kohd/

Belonging to the family of forward error correction techniques, this channel code feeds an information bit stream through a shift register so that each group of output bits depends on the current input and a span of previous inputs. Unlike a block code, encoding is continuous and stateful, which is why decoding relies on the maximum-likelihood Viterbi decoder walking a trellis. The encoder is defined by its constraint length K, its code rate R = k/n, and a set of generator polynomials; the canonical rate-1/2, K=7 code (generators 171, 133 octal) yields a free distance d_free = 10 and about 5.2 dB of soft-decision coding gain at a BER of 10⁻⁵. Convolutional codes are foundational in IEEE 802.11a/g, GSM, CCSDS deep-space links, and DVB satellite transmission.

Category: Signal Processing

Canonical Code: R=1/2, K=7

Free Distance: d_free = 10

How a Convolutional Encoder Builds Redundancy

The encoder is a finite-state machine built from a shift register of K−1 memory stages and a bank of modulo-2 adders. As each information bit shifts in, the adders combine taps from the register according to fixed generator polynomials to produce n output bits per k input bits. Because the present output depends on a window of recent inputs, the code carries memory; the constraint length K quantifies how many bit positions, including the current one, influence each output group. A rate-1/2, K=7 encoder therefore has six memory stages and 2⁶ = 64 states, and its behavior is captured completely by a state diagram or, unrolled over time, a trellis.

Decoding inverts this process by searching the trellis for the path whose coded sequence is closest to the received samples. The Viterbi algorithm performs this maximum-likelihood search efficiently using add-compare-select recursion, retaining only the survivor path into each state. Soft-decision metrics, where the demodulator passes quantized confidence values rather than hard 0/1 bits, recover roughly 2 dB more coding gain than hard-decision metrics. The decoder must accumulate a traceback depth of about five times the constraint length (35 trellis steps for K=7) before the survivor paths merge reliably and a decision bit can be released.

Throughput can be raised without redesigning the encoder by puncturing: the transmitter periodically deletes selected coded bits to convert the native rate-1/2 stream into higher rates such as 2/3, 3/4, 5/6, or 7/8. The decoder inserts erasures at the punctured positions and runs the same Viterbi machine, trading a fraction of a dB to several dB of coding gain for additional payload capacity. This is exactly how 802.11a adapts coding rate across its modulation-and-coding-scheme table.

Governing Relationships

Number of trellis states:
N_states = 2^(K−1)  (K = constraint length)

Code rate and bandwidth expansion:
R = k / n  →  BW expansion = 1 / R

Asymptotic coding gain (soft decision, AWGN):
G_a ≈ 10 × log₁₀(R × d_free) dB

Encoder output (rate 1/2 example):
v₁[n] = ∑ g_1,i · u[n−i]  (mod 2),  v₂[n] = ∑ g_2,i · u[n−i]  (mod 2)

Where u = input bit stream, g = generator polynomial taps, d_free = minimum free distance. Example: R=1/2, K=7, d_free=10 → G_a ≈ 10×log₁₀(5) ≈ 7 dB asymptotic, about 5.2 dB realized at BER 10⁻⁵.

Common Convolutional Codes and Their Parameters

Code (octal generators)	Rate R	Constraint K	d_free	Soft gain @ 10⁻⁵	Typical Use
5, 7	1/2	3	5	~3.5 dB	Low-complexity links
171, 133	1/2	7	10	~5.2 dB	802.11a/g, CCSDS, GSM
171, 133 (punctured)	3/4	7	4	~3.7 dB	Higher-throughput Wi-Fi
561, 753	1/2	9	12	~5.7 dB	Deep-space, NASA
171, 133, 165	1/3	7	15	~5.8 dB	IS-95, CCSDS rate-1/3

Common Questions

Frequently Asked Questions

What do constraint length and code rate mean for a convolutional code?

Constraint length K is the number of bit positions, including the current input, that affect each output group; it equals the shift-register length plus one, so K=7 means 6 stages and 2⁶ = 64 trellis states. Code rate R = k/n is input bits per output bits, so rate 1/2 doubles the symbol stream. Larger K raises d_free and coding gain but Viterbi complexity grows as 2^(K−1), so practical codes use K = 5 to 9. The canonical 802.11a/CCSDS code is rate-1/2, K=7 with generators 171 and 133 octal.

How does a convolutional code differ from a block code like Reed-Solomon?

A block code adds parity to fixed-length blocks independently and has no memory between blocks; a convolutional code carries memory in its shift register so output depends on a running window of recent inputs. Convolutional codes pair naturally with soft-decision Viterbi decoding and excel against random AWGN errors, while Reed-Solomon excels against bursts. Concatenated systems use an outer Reed-Solomon (255,223) around an inner convolutional code, the classic deep-space and DVB arrangement.

What coding gain does a rate-1/2 K=7 code give at a BER of 10⁻⁵?

With soft-decision Viterbi decoding on an AWGN channel it provides about 5.2 dB of gain, cutting required E_b/N₀ from roughly 9.6 dB (uncoded BPSK) to about 4.4 dB. Hard-decision decoding loses roughly 2 dB. Puncturing to rate 3/4 trades about 1.5 dB of gain for throughput. The asymptotic limit is near 10×log₁₀(R×d_free); real systems fall short at moderate BER due to error-event multiplicity.

Related Terms

← Convolution Convolutional Interleaver →